Paper2.txt

Perception and Preference Analysis of Fashion Colors

1. OVERVIEW
Color is light carried on wavelengths absorbed by the eyes that the brain converts into colors that we see, and it is ubiquitous and is a source of information [1]. Colors can have a powerful psychological effect, and there is a strong connection between color and feelings. Color can evoke emotions and therefore it can change our behavior too (a red sports car can create feelings of excitement, or a blue sea can create feelings of calmness). This is also supported by science, as color addresses one of our basic neurological needs for stimulation [2]. Color triggers very specific responses in the brain and in the whole body (red raises blood pressure and heart rates, while blue lowers blood pressure, pulse, and respiration rates). 

Color, as an important feature of the products and environments, can make great influences on the consumers’ preference and purchase intention, such as clothing, car interior, phone shells, furniture and house. Especially, Color, Materials, Finish (CMF) are the focuses in an area of industrial design as the chromatic, tactile and decorative identity of products and environments. Product colors make a bigger impact than we can imagine. From influencing customers’ purchase decisions to affecting customers’ behaviors, colors can also have vast implications on people’s ability to read, learn and comprehend. The prudent use of colors can contribute not only to differentiating products from competitors, but also to influencing moods and feelings.

For consumers, the visual appearance and color is placed above other factors when shopping (1% sound / smell, 6% texture, 93% visual appearance) [3]. The consumers make up their minds within 90 seconds of their initial interactions with either people or products and about 62~90% of the assessment is based on colors alone [1].

For marketers, color differentiates the brand, suggests emotional benefits and can be a key to a brand’s identity. Brand recognition directly links to consumer confidence and color increases brand recognition by 80%. [3]. Through color, a brand can establish an effective visual identity, form strong relationships with a target market, and position itself among competitors in the marketplace [4]. 

For retailers, shopping is the art of persuasion. Though there are many factors that influence how and what consumers buy. 85% of shoppers place color as a primary reason for why they buy a particular product [3]. However, a great deal is decided by visual cues, the strongest and most persuasive color.

To sum up, the perception of a color directly affects the parts of person’s nervous system to arouse various emotions, such as excitement, energy, and calmness, which play an important role for customers in making decisions on what they like and dislike. On another hand, to know the consumers’ preference of product color can help companies to reduce the stock, improve the sale and enhance the competitiveness. Hence, given that our moods and feelings are unstable and that colors play roles in forming attitude, it is significant to investigate the perception and preference of product color.

Furthermore, there is a significant difference between the preference between genders when coming to color selection. The study was done for most favorable and least favorable colors and there was a significant favor for the blue color of both men and women and the orange color was the most disliked color by both men and women [5]. Hence, it is also significant to investigate the preference difference of product color between female and male.

Generally, different combinations of the color, materials and finish of the product can affect people’s perception and preference. Furthermore, people have different perception and preference even towards the same colors of different products. In this proposal, we select the shirts as the main research objects to establish the model of people’s perception and preference on fashion, and analyze the aesthetic differences and similarities between male and female. Contrary to previous approaches, by using our results of a psychological experiment, we can build a layered model, which provides insights into hierarchical relationships involved in human aesthetic color perception (sometimes also taking pattern and texture into consideration). A hypothetical hierarchical feed-forward model of aesthetic color perception is shown in Figure 1. This model can use a set of intermediate judgements to link computational color features with aesthetic color properties. Our objective is to find the relationship among the intermediate aesthetic properties and finally confirm structure of the hierarchical feed-forward model of aesthetic perception for fashion colors.
 
Figure. 1. Structure of the hypothetical hierarchical feed-forward model of aesthetic perception for fashion colors. The model consists of 5 layers, containing 17 intermediate aesthetic properties.

Clothing has no gender; it is classified into men's and women's clothing because of the gender of the wearers. It can be seen that the gender of clothing is a consensus established by society. No other commodity in the world has such distinct gender differences as clothing. Compared with other products, the shirts are a kind of common products and popular fashion among male and female, including solid color, multi-color plaid shirts and printing shirts with color gradient, shown in Figure 2. Hence, it was reasonable to consider clothing rather than other products as the research object. 
     
(a)             (b)             (c)             (d)            (e)
Figure 2. Examples of various color shirts. (a) Solid color. (b) Two-color combination. (c) Multi-color combination. (d) One-color gradient. (b) Two-color gradient.

To build the hierarchical feed-forward model of aesthetic perception for fashion colors, it is significant to analyze their inherent features. The color schemes should be designed by combing with the exiting popular colors before doing the psychological experiment. Hence, the investigation method to analyze people’s perception and preference for fashion colors consists of three parts: 

1. Perception and preference analysis of qualitative colors (consists of solid colors, two-color combinations, multi-color combinations (more than two colors)), one-color gradient and two-color gradient. Colors have inherent features, like cold-warm, heavy-light and passive-active. The colors on the surface of different products and environments have different additional features, like masculine-feminine, hard-to-match-easy-to-match, traditional-fashionable, technological- untechnological. It is significant to analyze the perception and preference of qualitative colors just based on their inherent features.
This experiment consists of four steps. Firstly, we select several aesthetic properties of colors as the semantic scales to evaluate the inherent features of colors. Secondly, we select different discrete colors that are distributed uniformly in CIE L*C*h color space as the evaluation objects. Thirdly, the color patches are assessed by a certain number of observers (the number of males is nearly equal to that of females) in terms of the semantic scales. Finally, the relationships between the color values in CIE L*a*b and CIE L*C*h color space and assessed values of semantic scales are analyzed by data analysis and visualization methods. Additionally, the relationship of observers’ preference between the patch colors and object colors will be analyzed further in the third part. 

2. Preference analysis of colors of online existing shirts to obtain the popular solid colors and multicolor combinations. There are large-scale clothing images and their corresponding sales on different online selling platform, such as American Amazon, Chinese Taobao and Jingdong. To our best knowledge, it is the first attempt to collect and analyze colors from online large-scale clothing images automatically. Firstly, the web crawler technology is utilized to collect large-scale shirt images and sales from online selling platform. Secondly, the deep learning and image processing methods are used to extract the colors from the clothing images automatically. Thirdly, clustering algorithm is adopted to classify the colors and extract the popular solid colors and multi-color combinations for the male and female. The selected popular colors can be referred further when designing the color schemes in the next part.  

3. Perception and preference analysis of various color shirts, and investigation of their aesthetic differences and similarities between male and female, including solid color, multi-color plaid shirts and printing shirts with color gradient. Firstly, several aesthetic properties of colors are selected as the semantic scales to evaluate the clothing colors, such as masculine-feminine, hard-to-match-easy-to-match and traditional-fashionable. Secondly, combined with the former selected popular colors, different colors that are distributed uniformly in CIE L*C*h color space are selected as the evaluation objects. Thirdly, the fabrics and shirts are simulated successively based on the color combinations and fabric pattern. Fourthly, the stimulated shirts are assessed by a certain number of observers (the number of males is nearly equal to that of females) in terms of the semantic scales. Finally, the relationships between the color values in CIE L*a*b and CIE L*C*h color space and assessed values of semantic scales are analyzed by data analysis and visualization methods to investigate their aesthetic differences and similarities between male and female, and the relationships of assessed values between color blocks and shirts are also analyzed to investigate the aesthetic differences and similarities between color blocks and product color. 

Additionally, based on the established model for color shirts, we will validate and improve it by using other popular clothing, such as T-shirt, down jacket and sweater, to find the relationship among the intermediate aesthetic properties and finally confirm structure of the hierarchical feed-forward model of aesthetic perception for fashion colors. Besides clothing, we also will extend it to other, such as the vacuum cup, home and office walls, phone shells. 

However, the fashion colors are our main research object, but not our unique research object. The other popular industrial products and environmental colors also will be considered, sometimes by combing the product materials and finish, or using the advanced virtual reality.
?
2. RESEARCH BACKGROUND 
The chromatic, tactile and decorative identity of products and environments are the focuses in an area of industrial design: Color, Materials, Finish (CMF). Especially, Color is an important feature of the products and environments that can make great influences on the consumers’ preference and purchase intention, such as clothing, car interior, phone shells, furniture and house. Furthermore, to know the consumers’ preference of product color can help the companies to reduce the stock, improve the sale and enhance the competitiveness. Basically, the perception of a color directly affects the parts of person’s nervous system to arouse various emotions, such as excitement, energy, and calmness, which play an important role for customers in making decisions on what they like and dislike. Hence, it is significant to investigate the perception and preference of product color.

(1) General context perception and preference analysis
The ability to appreciate the aesthetic qualities of the items is one of the universal characteristics of humanity, including the colors, textures and pattern. Neuroaesthetics is the discipline of investigating how beauty activates aesthetic perception and appreciation. It is situated at the intersection of psychological aesthetics, neuroscience, and human evolution [6]. The main objective of neuroaesthetics is to “characterize the neurobiological foundations and evolutionary history of the cognitive and affective processes involved in aesthetic experiences and artistic and other creative activities” [7]. It has taken nearly two decades for neuroaesthetics to establish itself as a serious discipline concerned with the scientific investigation of aesthetics from a neurobiological perspective [8]. Louise P. Kirsch et al. summarized the brain systems involved in aesthetic perception of fine arts and human body in the review, which is shown in Figure 2. There are 27 brain regions involved in aesthetics appreciation of visual art. To some extent, the brain regions powerfully support that aesthetic perception, appreciation, and judgement have neural foundations [9].

  
Figure. 2. Schematic representation of the neural circuits implicated in aesthetic judgement tasks. Reprinted from the article published byKirsch et al.[10].  : Frontal and Reward areas;   : Sensorimotor areas;   : Visual areas. In blue, brain regions associated with reward processing, OFC = orbitofrontal cortices, vmPFC = ventromedian prefrontal cortex, ACC = anterior cingulate, AMG = amygdala; aI = anterior insula, and NAcc = nucleus accubens; in red, sensorimotor areas, M1 = primary motor area, S1 = primary somatosensory area, IPL = inferior parietal lobule, PMC = premotor cortex; in orange, visual areas, part of the occipitotemporal cortex: EBA= extrastriate body area, MT = motion integration area, EV= early visual area, PPA= parahippocampal place area, and pSTS = posterior superior temporal sulcus.

Most previous neuroaesthetic studies have focused on the investigation of the neural mechanisms underlying aesthetic appreciation as well as those factors that make certain stimuli, such as the general visual arts, dance, the human face, and music. Over the past decade, several neutrally based and neutrally inspired models of aesthetic perception have been proposed, such as The Neuropsychological Model [11], The Information-Processing Model [12, 13], The Mirror Model of Art [14], The Quartet Model of Human Emotion [15], The Unifying Model of Visual Aesthetic Experience [16] and The Hierarchical “Feed-Forward” Model [17, 18]. 
Among them, the first real mathematical model to bridge the gap between low-level statistical features and aesthetic emotions aroused by visual textures is proposed by Thumfart and colleagues [17]. Using the results of a psychological experiment, they modeled the relationship between computational texture features and their aesthetic properties. In contrast to previous approaches, this layered model provides insight into the hierarchical relationships involved in the aesthetic experience of texture properties. The structure of the hierarchical feed-forward model of aesthetic texture perception consists of three layers: Affective layer (How can the texture be described), Judgment layer (What the object says about itself) and Emotional layer (What do I feel when interacting with the texture). 
However, its simple internal structure is useful for controlling algorithm complexity and structural risk. Nonetheless, the development of a universal model of human aesthetic perception, appreciation, and judgement is underway in neurasthenics research [19].

(2) Product color perception and preference analysis
Besides the textures and patterns of the products and environments, during the color perception process, an associate feeling or emotion is normally induced in our brains. The term color emotion was used recently by researchers in this field to represent this feeling or emotion. Many researchers have suggested that color directly affects the parts of human’s nervous system that are responsible for emotion arousal, and different color s or color combinations usually have different meanings for people. As color emotion is in the domain of psychology, it is influenced by many factors such as sex, age, climate and geographic conditions, as well as race and cultural influences. 
The series of L. C Ou et al. studies clarify the relationship between color emotion and color preference in terms of solid colors and two-color combinations [20-22]. The study of J. H. Xin et al. investigates and compares the color emotions among three different regions and find out the direct relationship between color emotions of subjects from these three regions with the use of color planners [23, 24]. A. C. Hurlbert et al. report a robust, cross-cultural sex difference in color preference [25]. L. Beke et al. and L. C Ou et al. investigate the differences between young and aged observers’ color image preference respectively [26, 27]. Won S and Westland S concerned with whether colour meanings are affected by context [28]. Note that most of color preference studies were based on color patches. As pointed out by Norman and Scott, [29] such studies did not measure the observer responses to colors, but to color patches. It is still unclear whether the observer responses to color patches can generalize to real-world applications.

Colors have inherent features, like cold-warm, heavy-light and passive-active. The colors on the surface of different products and environments have different additional features, like masculine-feminine, hard-to-match-easy-to-match, traditional-fashionable, technological and untechnological. Schloss et al.[30] study object color preference using colored images of various objects, including couch, T-shirt, walls and two types of cars. As a result, color preference ratings are found to vary across different objects, suggesting a strong impact of context on color preference. Schloss et al. argue that for a particular object, observers tended to like colors that were “appropriate” for that object. Nevertheless, as Schloss et al. also mentioned, which colors are “appropriate” depended on the object, implying that the term “appropriate” was not well-defined.

	There were only a few studies of color preference that used colored objects or environments as the stimuli, such as cars [31], textile or fabrics [32-34], residences [35], room interiors[36]. Generally, different combinations of the color, materials and finish of the product can affect people’s perception and preference. Furthermore, people have different perception and preference even towards the same colors of different products, because the context of the products can affect colour perception and preference. Hence, it is hard to establish one model that can predict people’s perception and preference towards colors of different products. But, we can propose a universal investigation method that can be used to find the corresponding pattern and establish the corresponding model of people’s perception and preference towards colors of various products. Using existing research findings of “colour emotion”[20-24] and colour harmony[27, 37, 38] as a theoretical basis, this study aims to discover factors affecting object colour perception and preference.

The main object investigated in this proposal is to find the relationship among the intermediate aesthetic properties and confirm structure of the hierarchical feed-forward model of aesthetic perception for fashion colors, especially in terms of the aesthetic differences and similarities between male and female，finally use other industrial products and environmental colors to validate and improve the proposal model. Compared with other products, the shirts are a kind of common products and popular clothing among male and female, including solid color, multi-color plaid shirts and printing shirts with color gradient, shown in Figure 2. This project proposes to address the challenges of:  
Challenge 1: Perception and preference analysis of qualitative colors (consists of solid colors, two-color combinations, multi-color combinations (more than two colors)), one-color gradient and two-color gradient. To our best knowledge, it is the first attempt to investigate the perception and preference of color gradient, although there is much research on the perception and preference of solid colors and two-color combinations [20-25]. The color gradient is applied into product design and there is still no related research about their perception and preference analysis. Additionally, the perception and preference analysis of multi-color combinations is always a hard problem, because it is harder to describe an effective features of color combinations along with the increasing number of colors.

Challenge 2: Preference analysis of colors of online existing shirts to obtain the popular solid colors and multicolor combinations. To our best knowledge, it is the first attempt to obtain the popular solid colors and multicolor combinations by extracting and analyzing the colors from large-scale online shirt images with the help of deep learning methods [39-43]. With the development of online shopping, there are a large-scale clothing images on various online shopping platform. But it is hard to recognize and segment the object from the images automatically until the novel deep learning methods are proposed recently. Hence, we can extract colors from large-scale online shirt images and analyze their popular colors by using computer vision and deep learning methods. 

Challenge 3: Perception and preference analysis of various color shirts, and investigation of their aesthetic differences and similarities between male and female, including solid color, multi-color plaid shirts and printing shirts with color gradient. Because the previous studies do not apply the 3D clothing simulation technology, they just do some research on the investigation of 2D fabric color preference [33, 34]. But there is a difference of color preference between 2D fabric and 3D clothing. Hence, to our best knowledge, it is also the first attempt to investigate the color preference of 3D clothing. Furthermore, most previous neuroaesthetic studies have focused on the investigation of the neural mechanisms underlying aesthetic appreciation as well as those factors that make certain stimuli [11-18]. However, these proposed models is used for general context perception and preference analysis, rather than the specific object whose colors are their main features. Hence, we still need to establish an effective model to bridge the gap between image colors and aesthetic emotions aroused by visual colors.  

Challenge 4: Extension of investigation method of perception and preference of shirt colors. The investigation method of perception and preference of shirt colors are applied into other objects. There were only a few studies of color preference that used colored objects or environments as the stimuli [31-36], we still need to apply the proposed investigation method of perception and preference of product colors into other popular and common products, such as vacuum cup, home and office walls, phone shells. Furthermore, we should utilize the advanced simulation technology to make the simulated products and virtual environment more real and reliable, thereby obtaining the more real perception and preference data from observers, such as the advanced virtual reality (VR) [44].
?
RESEARCH METHODOLGY
In our proposal, we will use data analysis and visualization, machine learning, computer vision and even deep learning method to investigate people’s perception and preference of product and environment colors. The investigation method to analyze people’s perception and preference consists of four main parts: 
(1) Perception and preference analysis of qualitative colors (consists of solid colors, two-color combinations, multi-color combinations (more than two colors)), one-color gradient and two-color gradient. Colors have inherent features, like cold-warm, heavy-light and passive-active. The colors on the surface of different products and environments have different additional features, like masculine-feminine, hard-to-match-easy-to-match, traditional-fashionable, technological and untechnological. It is significant to analyze the perception and preference of qualitative colors just based on their inherent features.
This experiment consists of four steps. Firstly, we select several aesthetic properties of colors as the semantic scales to evaluate the inherent features of colors. Secondly, we select different discrete colors that are distributed uniformly in CIE L*C*h color space as the evaluation objects. Thirdly, the color patches are assessed by a certain number of observers (the number of males is nearly equal to that of females) in terms of the semantic scales. Finally, the relationships between the color values in CIE L*a*b and CIE L*C*h color space and assessed values of semantic scales are analyzed by data analysis and visualization methods. Additionally, the relationship of observers’ preference between the patch colors and object colors will be analyzed further in the third part. 

(2) Preference analysis of colors of online existing shirts to obtain the popular solid colors and multicolor combinations. There are large-scale clothing images and their corresponding sales on different online selling platform, such as American Amazon, Chinese Taobao and Jingdong. Firstly, the web crawler technology is utilized to collect large-scale shirt images and sales from online selling platform. Secondly, the deep learning and image processing methods are used to extract the colors from the clothing images automatically. Thirdly, clustering algorithm is adopted to classify the colors and extract the popular solid colors and multi-color combinations for the male and female. The selected popular colors can be referred further when designing the color schemes in the next part.  

3. Perception and preference analysis of various color shirts, and investigation of their aesthetic differences and similarities between male and female, including solid color, multi-color plaid shirts and printing shirts with color gradient. Firstly, several aesthetic properties of colors are selected as the semantic scales to evaluate the clothing colors, such as masculine-feminine, hard-to-match-easy-to-match and traditional-fashionable. Secondly, based on the former selected popular colors, different colors that are distributed uniformly in CIE L*C*h color space are selected as the evaluation objects. Thirdly, the fabrics and shirts are simulated successively based on the color combinations and fabric pattern. Fourthly, the stimulated shirts are assessed by a certain number of observers (the number of males is nearly equal to that of females) in terms of the semantic scales. Finally, the relationships between the color values in CIE L*a*b and CIE L*C*h color space and assessed values of semantic scales are analyzed by data analysis and visualization methods to investigate their aesthetic differences and similarities between male and female, and the relationships of assessed values between color blocks and shirts are also analyzed to investigate the aesthetic differences and similarities between color blocks and product color. 

(4) Extension of investigation method of perception and preference of shirt colors. The investigation method of perception and preference of shirt colors are applied into other objects to test its universality, including the vacuum cup, home and office walls, phone shells. (1) Vacuum cups are usually made of stainless steel, purple sands, ceramics, glass or plastic. It is significant to investigate the perception and preference of combinations of solid colors and materials for vacuum cups, which focus on the effect of heat preservation. (2) Phone shells with color gradient made by various phone brands came into being in China recently and become more and more popular, which focus on the “technological” and “fashionable”. (3) Colors of home and office walls is the key of the environmental atmosphere, which focus the “warm and sweet” and “official and passionate” respectively. 

1.	Perception and preference analysis of colors.
The popular colors include qualitative colors (consists of solid colors, two-color combinations, multi-color combinations (more than two colors)), one-color gradient and two-color gradient, which usually applied into designing products, shown in Table 1. 
In computer graphics, a color gradient (sometimes called a color ramp or color progression) specifies a range of position-dependent colors, usually used to fill a region, including one-color, two-color and multi-color gradient. An axial color gradient (sometimes also called a linear color gradient) is specified by two points, and a color at each point. The colors along the line through those points are calculated using linear interpolation, then extended perpendicular to that line. In digital imaging systems, colors are typically interpolated in an RGB color space, often using gamma compressed RGB color values, as opposed to linear [45]. 
Table 1. Color types
Type	Example 1	Example 2	Example 3
Solid color	 	 	 
Two-color combination	 	 	 
Multi-color combination	 	 	 
One-color gradient	 	 	 
Two-color gradient	 	 	 
Multi-color gradient	 	 	 

1.1 Solid color and one-color gradient
	The series of L. C Ou et al. studies clarify the relationship between color emotion and color preference in terms of solid colors and two-color combinations [20-22]. The performance of solid-colour-emotion models developed by L. C Ou et al.[20] for predicting the present experimental data is shown in Table 2. In Table1, L* is CIELAB lightness, C* is CIELAB chroma, h is CIELAB hue angle, ΔC*N5 is the CIELAB chroma difference between a test color and a medium gray with L* of 50, and ΔL*N5 is CIELAB lightness difference between a test color and a medium gray with L* of 50.
Table 2. The performance of solid-colour-emotion models developed by L. C Ou et al.[20] for predicting the present experimental data.
Emotion	Models 	R2
Warm-Cool	WC=-0.5+0.02(C*)1.07cos(h-50o)	0.74
Heavy-Light	HL=-2.1+0.05(100-L*)	0.76
Active-Passive	 
0.75

	They just find out the relationship between color emotions and color values in CIELAB color space in terms of solid colors (shown in Table 2), but not color combinations. Besides, they do not do any research on the perception and preference analysis of color gradients. Therefore, we still need to do more research on the relationship between color emotion and color preference in terms of color combinations and color gradients. 	 
We can take the one-color gradient as an example to illustrate the investigation process. This experiment consists of four steps. Firstly, we select several aesthetic properties of colors as the semantic scales to evaluate the inherent features of colors, including Warm-Cool, Heavy-Light, Active-Passive, and Like-Dislike. Secondly, we select different discrete one-color gradient that are distributed uniformly in CIE L*C*h color space as the evaluation objects. Thirdly, the color patches are assessed by a certain number of observers (the number of males is nearly equal to that of females) in terms of the semantic scales. Finally, the relationships between the color values in CIE L*a*b and CIE L*C*h color space and assessed values of semantic scales are analyzed by data analysis and visualization methods. Additionally, the relationship of color emotions between color gradient and solid color are also analyzed.
    
Figure 3. 36 Solid Colors distributed in CIE L*a*b color space.

1.2 Two-color combinations and gradient
In terms of two-combinations, the combinations of white (or black) and other colors are common in our daily life, such as clothes. We have used the Internet to collect pictures of 190 pieces of two-color plaid shirts from the eight brands (H&M, Zara, Uniqlo, C&A, GAP, Muji, Forever 21, and New Look). The selection criteria included: a) two-color plaids, and b) shirts, regardless of material. Of the total, 117 pieces were men's shirts, while the other 73 were women's. The colors in the pictures of these shirts were extracted by a computer. In line with the CIE L* values (Li,1, Li,2) under the CIE LAB color space, the 117 pairs of colors were classified into 117 light colors and 117 dark colors, in which i indicates the i-th shirt, and 1 and 2 stand for Color 1 and Color 2 of the i -th shirt, as shown in Figure 4. 
 
(a)	 
(b)
Figure 4. Colors distributions of two-color combinations. 
(a) the lighter color. (b) the darker color.
It could be seen from the above figure that black and white were used the most in the two-color plaid shirts, and that black and white was the most common colors among the color combinations. Therefore, we firstly analyze the perception and preference of combinations of white (or black) and other colors. Then we will do some research on multi-color combinations and gradients further.   

2.	Preference analysis of colors of online existing shirts
There are large-scale clothing images and their corresponding sales on different online selling platform, such as American Amazon, Chinese Taobao and Jingdong. We can take full advantages of these online clothing images to extract their colors automatically and analyze their popularity. 
Firstly, the web crawler technology is utilized to collect large-scale shirt images and sales from online selling platform (some online sample images are shown in Figure 5(a)). Secondly, the deep learning and image processing methods are used to extract the colors from the clothing images automatically. Before clothing color extraction, it is important to detect the clothing object and estimate the keypotins (shown in Figure 5(b)) to segment the clothing from clothing images automatically. Thirdly, clustering algorithm is adopted to classify the colors and extract the popular solid colors and multi-color combinations for the male and female. 
     
(a)                            (b)                    (c)
Figure 5. Online clothing samples (a) and example with the keypoints (b) and annotations (c).


It is shown in Figure 5 that the flow of automatic keypoints recognition and color extraction to process independent clothing image. In Figure 5, it is a top-down pipeline. The Stage 1 consists of   Detecting and   Cropping the object in order to remove the unrelated contexts in the clothing image and detect the multiple models in one image, which can be realized by Faster R-CNN [40]. The Stage 2 is   localizing the keypoints of clothing, which can be realized by Stacked Hourglass Network [41]. The stage 3 is to use Mean shift clustering algorithm for   segmenting the clothing and extract its colors
 
Figure 5. A top-down pipeline of automatic keypoints recognition and color extraction.

2.1	Automatic Keypoints Recognition 
Recently, deep representation learning has been successfully applied to various computer vision areas, such as image classification [39, 46, 47] , object detection [40, 42], human pose estimation [41, 48, 49] and clothing parsing [50, 51]. We can build our feature extraction model and realize automatic keypoints recognition based on Faster R-CNN [40] (region-based convolutional neural network) (to detect the clothing object) and Stacked Hourglass Network [41] (to estimate the visible keypoints of clothing, as shown in Figure 5.
Faster R-CNN. Faster R-CNN is one of the State-of-the-art object detection networks. It consists of three parts: the very deep VGG-16 model to extract feature, Region Proposal Network (RPN) to enable nearly cost-free region proposals, classification and regression [40]. We can adopt the Faster R-CNN to detect the clothing object in the image.
Stacked Hourglass Network. The Stacked Hourglass Network is proposed to determine the precise pixel location of important keypoints of the body and State-of-the-art results are achieved on the FLIC and MPII benchmarks outcompeting all recent methods. On MPII there is over a 2% average accuracy improvement across all joints, with as much as a 4-5% improvement on more difficult joints like the knees and ankles [41]. The clothing keypoints is similar to the body keypoints. Therefore, we can apply it into estimating the keypoints of clothing after using the data of clothing with keypoints and annotations to train the Stacked Hourglass Network. 
The Stacked Hourglass Network consists of multiple stacked hourglass modules which allow for repeated bottom-up, top-down inference, as shown in Figure 6. The network captures and consolidates information across all scales of the image. The design is referred as an hourglass based on our visualization of the steps of pooling and subsequent upsampling used to get the final output of the network. Like many convolutional approaches that produce pixel-wise outputs, the hourglass network pools down to a very low resolution, then upsamples and combines features across multiple resolutions. On the other hand, the hourglass differs from prior designs primarily in its more symmetric topology. A single hourglass is expanded by consecutively placing multiple hourglass modules together end-to-end. This allows for repeated bottom-up, top-down inference across scales. In conjunction with the use of intermediate supervision, repeated bidirectional inference is critical to the network’s final performance.
 
Figure 6. The network for pose estimation consists of multiple stacked hourglass modules which allow for repeated bottom-up, top-down inference [41].
The full module (excluding the final 1×1 layers) is illustrated in Figure 7. The topology of the hourglass is symmetric, so for every layer present on the way down there is a corresponding layer going up. After reaching the output resolution of the network, two consecutive rounds of 1×1 convolutions are applied to produce the final network predictions. The output of the network is a set of heatmaps where for a given heatmap the network predicts the probability of a joint’s presence at each and every pixel. 
 
Figure 7. An illustration of a single “hourglass” module. Each box in the figure corresponds to a residual module. The number of features is consistent across the whole hourglass [41].

2.2	Clothing Color Extraction
We have already applied Mean shift clustering algorithm into extract colors form Chinese traditional costumes automatically. The experimental results demonstrate that the proposed method can extract the dominant colors from costumes images with great accuracy when the bandwidth of Mean shift clustering algorithm is set as 0.05 [52]. The Mean shift clustering algorithm also can be used to extract the colors of upper-body clothes.
	The Mean shift clustering algorithm is a nonparametric iteration method based on the increased gradient density [53]. It is a hill climbing algorithm which involves shifting a kernel iteratively to a higher density region until convergence. Every shift is defined by a mean shift vector, which always points toward the direction of the maximum increase in the density. At every iteration the kernel is shifted to the centroid or the mean of the points within it. The method of calculating this mean depends on the choice of the kernel. The outstanding advantages of Mean shift clustering algorithm are small computations, which are simple and easy to realize. It can significantly reduce the number of basic image entities, and because of the good discontinuity preservation filtering characteristic, the salient features of the image are retained. 
Assume a circular window center at C and having radius h as the kernel. Given a set of N data points xi, i=1, 2,…, N, and xi?Ω, the probability density of x can be obtained from the kernel density estimator:
	 	(1)
The kernel K(x-xi) is generally given by a uniform function or a Gaussian function. The Gaussian function is as follows:
	 	(2)
Then the gradient of the density function becomes:
	 	(3)
Where g(||(xi -x)/h||2)=-k/(||(xi -x)/h||2), assuming that the derivative of the kernel profile k exists for all x ?[0, ∞)and h is the bandwidth to adjust the resolution for the difference between xi and x. The difference between the weighted mean and the center of the kernel is called “mean shift”. The mean shift algorithm moves iteratively each data point x in the feature space by the mean shift vector m(x) until the mean converges to an estimate of the local mode of the data set. The mean shift vector m(x) can be defined as:
	 	(4)
The mean shift clustering algorithm can be obtained based on the analysis in Equation (8). First, an image is represented as a two-dimensional (2D) lattice of p-dimensional vectors, where p>3 for multispectral images, p=3 for color images and p=1 for gray-level images. Here, it is needed to select the bandwidth parameter h which determines the resolution of mode detection by controlling the size of kernel. 

2.3	Color Preference Analysis 
After we obtain multitude of colors from large-scale clothing, we can use Mean sift clustering algorithm to classify the colors again, calculate the frequency and popularity of different colors and analyze their difference and similarity of preference for female and male. Here, we take the 480 colors extracted from solid shirts as an example to illustrate the color clustering and frequency analysis method.
Current solid shirt color analysis. We collect 480 colors from solid shirts of different clothing brands (shown in Figure 8(a)) and use Mean Shift clustering algorithm to classify them into 19 clusters in CIE Lab color space (shown in Figure 8(b)). 19 cluster centers and their corresponding numbers of clusters are shown in Figure 8(c), and the white is the most popular color, followed by the dark and red. 
   
(a)                      (b)                       (c)
Figure 8. Clustering results of shirts’ colors.
(a) Original colors. (b) Clustering results. (c) Clustering centers.

3.	Perception and preference analysis of various color shirts
Perception and preference analysis of various color shirts, and investigation of their aesthetic differences and similarities between male and female, including solid color, multi-color plaid shirts and printing shirts with color gradient. We take the solid color shirts as the example to illustrate the investigation method. In the beginning, we take the solid color shirts as the example to illustrate the perception and preference analysis method.
3.1	Solid color scheme designing 
Based on the 19 colors of cluster centers, another 22 solid colors are added to generate the solid color scheme, whose CIE Lab values are shown in Figure 9. The object is as possible as to make the selected solid colors be distributed uniformly in CIE Lab color space. 
 
Figure 9. Selected solid colors in CIE Lab color space.

3.2	Solid color clothing simulating 
We use MATLAB software to simulate the solid color fabric, as shown in Figure 10(a). Then, based on the simulated fabrics, CLO 3D software is used to simulate the three-dimensional men's and women's shirts, as shown in Figure 10(b) (in which the image on the left is a basic type of men's shirt for a height of 175cm, the middle is a basic type of women's shirt for a height of 165cm, and the right is a simulated drawing of fabric with a size of 15cm ? 15cm). It is shown in Figure 10 that 41 female shirts whose colors are corresponding to the selected colors in Figure 9. 
     
Figure 10. Simulated fabric and shirt. (a) Fabric. (b) Male shirt. (c) Female shirt.

                                         
Figure 11. Simulated female shirts.

3.3 Semantic scales designing
We collect multitude of semantic scales of aesthetic emotions from previous studies [18, 20, 23, 33] and select 10 pairs of semantic scales related to shirts: “Cold-Warm”, “Heavy: Light”, “Passive: Active”, “Tense: Relaxed”, “Plain: Splendid”, “Traditional : Modern”, “Masculine : Feminine”,  “Slim-Look: Fat-Look”, “Hard-to-Match: Easy-to-Match”, “Dislike: Like”. 
During the main experiment, each word pair was presented using a ten-step, forced choice scale to measure the observer response. Taking “Dislike: Like” as an example, the 10 steps included "extremely dislike", "exceptionally dislike", "fairly dislike", "dislike", "somewhat dislike", "somewhat like", "like", "fairly like", "exceptionally like", and "extremely like". The respondents are forced to make a choice between like/dislike rather than selecting a neutral response. A certain number of female and male observers are asked to fill in the same 10-step mandatory scale with choice questions to assess the 40 shirts.
In the experiment, this study can use two Apple 27-inch iMacs (model: MD095CH/A) as the display. The two monitors were under the same indoor environment and light source. During the experiment, other interference factors were avoided and the same brightness and display calibration are set for the two monitors. During the test, an image was displayed for 10 seconds to avoid interference with the next image. 

3.3	Results Analysis and Discussion
3.3.1	Variance analysis
Gender variance analysis. There is a significant difference between the preference between genders when coming to color selection. The study was done for most favorable and least favorable colors and there was a significant favor for the blue color of both men and women and the orange color was the most disliked color by both men and women [5]. Hence, it is also significant to investigate the preference difference of aesthetic emotions between male and female. Therefore, an independent sample t-test should be used to analyze the differences in the semantic scales for different solid color shirts from the viewpoint of gender.

Age variance analysis. A research Results indicate significant differences between young and aged observers’ color image preference, some of which can be explained with neuro-physio-logical changes, others may be attributed to cultural implications [26]. It is important to know the difference of aesthetic emotions among different age people. Therefore, an independent sample t-test is used to analyze the differences in the semantic scales for different solid color shirts from the viewpoint of age.

3.3.2 Regression analysis
It is possible and significant to quantify aesthetic emotion of solid color shirts and confirm the relationship between color values and aesthetic emotions. In our studies, the CIELAB values is used to describe the color of solid shirts.
The visible gamut plotted within the CIELAB color space is shown in Figure 12(a). The CIELAB color space (also known as CIE L*a*b* or sometimes abbreviated as simply "Lab" color space) is a color space defined by the International Commission on Illumination (CIE) in 1976. It expresses color as three numerical values, L* for the lightness and a* and b* for the green–red and blue–yellow color components. CIELAB was designed to be perceptually uniform with respect to human color vision, meaning that the same amount of numerical change in these values corresponds to about the same amount of visually perceived change.
The CIELCh color space is a CIELab cube color space, where instead of Cartesian coordinates a*, b*, the cylindrical coordinates C* (chroma, relative saturation) and h° (hue angle, angle of the hue in the CIELab color wheel) are specified, as shown in Figure 12(b). The CIELab lightness L* remains unchanged. The conversion of a* and b* to C* and h° is done using the following formulas:
	 	(5)
	 	(6)
     
(a)                                (b)
Figure 12. Color space. (a) The visible gamut plotted within the CIELAB color space. a and b are the horizontal axes. L is the vertical axis. Uses D65 whitepoint. (b) The visible gamut plotted within the CIELCH color space. L is the vertical axis; C is the radius; h is the angle around the circumference.

In our preliminary study, we asked 21 female and 21 male observers to fill in the same 10-step mandatory scale with choice questions to assess the 40 shirts. The relationship between CIE LCH values and aesthetic emotions should be analyzed and the fitting equations should be confirmed by regression analysis method, such as multiple linear regression, nonlinear regression, ridge regression, lasso Regression. The scatterplot matrix of aesthetic emotions is shown in Figure 13. The internal relationship can be analyzed. For example, the aesthetic emotion of “Dislike : Like” has a strong positive linear correlation with “Hard-to-Match: Easy-to-Match” whose correlation coefficient is 0.85.
 
Figure 13. The scatterplot matrix of aesthetic emotions. 
(Corr indicates the correlation coefficient, “FatL : SlimL” indicates Slim-Look: Fat-Look, “HMatch: EMatch” indicates Hard-to-Match: Easy-to-Match ) 

The internal relationship of aesthetic emotions can be analyzed. Based on the hierarchical feed-forward theory of aesthetic texture perception [17, 18], a hierarchical feed-forward model of aesthetic perception for solid color shirts can be developed in Figure 14. A hierarchical feed-forward model with five layers is established. However, the regression equations indicated by the linking lines with circle and arrow in Figure 14 should be confirmed by using least square method, and the hierarchical feed-forward model in Figure 14 are just suitable for solid color shirts. 
 
Figure 14. A hierarchical feed-forward model of aesthetic perception for solid color shirts.

Besides the solid color shirts, aesthetic perception for other color shirts also need to be investigated by the above similar flow, including two-color, multi-color, one-color gradient and two-color gradient shirts, as show in Figure 15. For these color shirts, some other aesthetic properties should be taken into consideration to the hierarchical feed-forward model, such as maximum and mean color difference in the inherent layer, Unharmonious: Harmonious, Simple: Complex in the affective layer, as shown in Figure 1. Compare with solid color shirts, the harmonious degree of the color combination and the complexity of the color pattern are important for two-color and multi-color shirts, and the color gradients of the one-color gradient shirts and the harmonious degree of the two-color gradient shirts are also significant aesthetic properties. 
Additionally, based on the established model for color shirts, we will validate and improve it by using other popular clothing, such as T-shirt, down jacket and sweater. Hence, a more sophisticated psychological experiment should be designed and implemented further to establish and improve a more comprehensive hierarchical feed-forward model of aesthetic perception for fashion colors to provide reference for fashion designers and sellers. 
     
(a)
     
(b)
     
(c)
     
(d)
Figure 15. Simulated shirts.
(a) Two-color. (c) Multi-Color. (c) One-color gradient. (d) Two-color gradient.

4. Extension application of investigation method.
Besides color shirts, there are still multitude of color products and environments whose colors make a great influence on their applications or sales, such as vacuum cups, phone shells and home and office walls. For another industrial product and environmental colors, some new aesthetic properties should be taken into consideration to the hierarchical feed-forward model, such as “technological: nontechnological” for phone shells, “warm: sweet” and “official: passionate” for home and office walls. We can use the above hierarchical feed-forward model and aesthetic perception investigation method to analyze these product and environment colors and find the consumers’ color preference. 

4.1 Phone shells: product colors 
Phone is a typical electronic product and phone shells with color gradient made by various phone brands came into being in China recently and become more and more popular, which focus on the “technological” and “fashionable”, as shown in Figure 18. It is important to analyze aesthetic emotions for people of different age and gender and realize personalized design to cater to different groups’ tastes.
 
Figure 16. Different phone shells with color gradient.

4.2 Home and office walls: environmental colors  
Colors of home and office walls is the key of the environmental atmosphere, which focus the “warm and sweet” and “official and passionate” respectively. Some different color walls at home are shown in Figure 19. Walls are the most important part of home and office that has the largest proportion of area. It is significant to analyze what emotions their colors can stimulate. 
We can even use the advanced virtual reality (VR) to simulate the home and office with different color furniture and walls. VR is an interactive computer-generated experience taking place within a simulated environment, that incorporates mainly auditory and visual, but also other types of sensory feedback like haptic. This immersive environment can be similar to the real world or it can be fantastical, creating an experience that is not possible in ordinary physical reality. Augmented reality systems may also be considered a form of VR that layers virtual information over a live camera feed into a headset or through a smartphone or tablet device giving the user the ability to view three-dimensional images [44].
          
Figure 17. Different color walls at home.

4.3 Vacuum cups: product materials and colors  
Vacuum cups are usually made of stainless steel, purple sands, ceramics, glass or plastic, as shown in Figure 16. It is significant to investigate the perception and preference of combinations of solid colors and materials for vacuum cups, which focus on the effect of heat preservation. Color, Materials, Finish (CMF) are the focuses is an emerging industrial design discipline within design as the chromatic, tactile and decorative identity of products and environments. Vacuum cups can used as a typical research object to investigate the comprehensive effect of colors and materials on consumers’ preference. 
  
Figure 18. Different solid color vacuum cups.


RESEARCH PLAN

Phase	Month (36)	Milestones
Perception and preference analysis of colors	6	Establish equations between aesthetic emotions and color values and qualify people’s color preference, consists of: 
?	Solid colors, two-color combinations, multi-color combinations (more than two colors), 
?	One-color gradient and two-color gradient.
Preference analysis of colors of online existing shirts	9	?	Collect large-scale shirt images and sales from online selling platform.
?	Realize automatic keypoints recognition by using deep learning method.
?	Realize automatic color extraction by using Mean Shift clustering algorithm.
?	Analyze people’s color preference and find the popular colors.
Perception and preference analysis of various color shirts	15	?	Analyze and cluster current shirt colors.
?	Design color scheme and semantic scales and simulate 3D color shirts.
?	Qualify people’s aesthetic emotions by machine learning method, such as regression analysis.
?	Establish a hierarchical feed-forward model of aesthetic perception by variance and correlation analysis. 
Extension application of investigation method	6	Use the above aesthetic perception investigation method to analyze these product and environment colors and find the consumers’ color preference:
?	Vacuum cups
?	Phone shells
?	Home and office walls


REFERENCE
1.	Singh, S., Impact of color on marketing. Management Decision, 2006. 44(6): p. 783-789.
2.	Hillenbrand, P., et al., Better branding: brand names can influence consumer choice. Journal of Product & Brand Management, 2013. 22(4): p. 300-308.
3.	Choi, H.H., S.A. Lim, and J. HeeKim, Promotional Video of Editing Techniques Utilizing Color and Brand Balance. International Journal of Software Engineering and Its Applications, 2014. 8(7): p. 149-158.
4.	Labrecque, L.I. and G.R. Milne, Exciting red and competent blue: the importance of color in marketing. Journal of the Academy of Marketing Science, 2012. 40(5): p. 711-727.
5.	Funk, D. and N. Oly Ndubisi, Colour and product choice: a study of gender roles. Management Research News, 2006. 29(1/2): p. 41-52.
6.	Bundgaard, P.F., Feeling, meaning, and intentionality-a critique of the neuroaesthetics of beauty. Phenomenology and the Cognitive Sciences, 2015. 14(4): p. 781-801.
7.	Chatterjee, A., Neuroaesthetics: A Coming of Age Story. Journal of Cognitive Neuroscience, 2011. 23(1): p. 53-62.
8.	Leder, H. and M. Nadal, Ten years of a model of aesthetic appreciation and aesthetic judgments: The aesthetic episode - Developments and challenges in empirical aesthetics. British Journal of Psychology, 2014. 105(4): p. 443-464.
9.	Boccia, M., et al., Where does brain neural activation in aesthetic responses to visual art occur? Meta-analytic evidence from neuroimaging studies. Neuroscience and Biobehavioral Reviews, 2016. 60: p. 65-71.
10.	Kirsch, L.P., C. Urgesi, and E.S. Cross, Shaping and reshaping the aesthetic brain: Emerging perspectives on the neurobiology of embodied aesthetics. Neuroscience and Biobehavioral Reviews, 2016. 62: p. 56-68.
11.	Chatterjee, A., Prospects for a cognitive neuroscience of visual aesthetics. 2003.
12.	Leder, H., Next Steps in Neuroaesthetics: Which Processes and Processing Stages to Study? Psychology of Aesthetics Creativity and the Arts, 2013. 7(1): p. 27-37.
13.	Leder, H., et al., A model of aesthetic appreciation and aesthetic judgments. British Journal of Psychology, 2004. 95: p. 489-508.
14.	Tinio, P.P.L., From Artistic Creation to Aesthetic Reception: The Mirror Model of Art. Psychology of Aesthetics Creativity and the Arts, 2013. 7(3): p. 265-275.
15.	Koelsch, S., et al., The quartet theory of human emotions: An integrative and neurofunctional model. Physics of Life Reviews, 2015. 13: p. 1-27.
16.	Redies, C., Combining universal beauty and cultural context in a unifying model of visual aesthetic experience. Frontiers in Human Neuroscience, 2015. 9.
17.	Thumfart, S., et al., Modeling Human Aesthetic Perception of Visual Textures. Acm Transactions on Applied Perception, 2011. 8(4).
18.	Liu, J.L., E. Lughofer, and X.Y. Zeng, Could linear model bridge the gap between low-level statistical features and aesthetic emotions of visual textures? Neurocomputing, 2015. 168: p. 947-960.
19.	Liu, J., E. Lughofer, and X. Zeng, Toward Model Building for Visual Aesthetic Perception. Computational intelligence and neuroscience, 2017. 2017.
20.	Ou, L.C., et al., A study of colour emotion and colour preference. Part I: Colour emotions for single colours. Color Research and Application, 2004. 29(3): p. 232-240.
21.	Ou, L.C., et al., A study of colour emotion and colour preference. part II: Colour emotions for two-colour combinations. Color Research and Application, 2004. 29(4): p. 292-298.
22.	Ou, L.C., et al., A study of colour emotion and colour preference. Part III: Colour preference Modeling. Color Research and Application, 2004. 29(5): p. 381-389.
23.	Xin, J.H., et al., Cross-regional comparison of colour emotions part I: Quantitative analysis. Color Research And Application, 2004. 29(6): p. 451-457.
24.	Xin, J., et al., Cross‐regional comparison of colour emotions Part II: Qualitative analysis. Color Research & Application, 2004. 29(6): p. 458-466.
25.	Hurlbert, A.C. and Y.Z. Ling, Biological components of sex differences in color preference. Current Biology, 2007. 17(16): p. R623-R625.
26.	Beke, L., et al., Color preference of aged observers compared to young observers. Color Research & Application: Endorsed by Inter‐Society Color Council, The Colour Group (Great Britain), Canadian Society for Color, Color Science Association of Japan, Dutch Society for the Study of Color, The Swedish Colour Centre Foundation, Colour Society of Australia, Centre Fran?ais de la Couleur, 2008. 33(5): p. 381-394.
27.	Ou, L.C., et al., Age effects on colour emotion, preference, and harmony. Color Research and Application, 2012. 37(2): p. 92-105.
28.	Won, S. and S. Westland, Colour meaning and context. Color Research and Application, 2017. 42(4): p. 450-459.
29.	Norman, R.D. and W.A. Scott, Color and affect: a review and semantic evaluation. The Journal of General Psychology, 1952. 46(2): p. 185-223.
30.	Schloss, K.B., E.D. Strauss, and S.E. Palmer, Object color preferences. Color Research and Application, 2013. 38(6): p. 393-411.
31.	Saito, T., Latent Spaces of Color Preference with and without a Context - Using the Shape of an Automobile as the Context. Color Research and Application, 1983. 8(2): p. 101-113.
32.	Kim, S.J., et al., Emotion-based textile indexing using colors, texture and patterns. Advances in Visual Computing, Pt 2, 2006. 4292: p. 9-18.
33.	Hsu, M.Y., L.C. Ou, and S.S. Guan, Colour preference for Taiwanese floral pattern fabrics. Color Research and Application, 2016. 41(1): p. 43-55.
34.	Kim, E.Y., et al., Emotion-based textile indexing using colors and texture. Fuzzy Systems and Knowledge Discovery, Pt 1, Proceedings, 2005. 3613: p. 1077-1080.
35.	Manav, B., Color‐emotion associations and color preferences: A case study for residences. Color Research & Application: Endorsed by Inter‐Society Color Council, The Colour Group (Great Britain), Canadian Society for Color, Color Science Association of Japan, Dutch Society for the Study of Color, The Swedish Colour Centre Foundation, Colour Society of Australia, Centre Fran?ais de la Couleur, 2007. 32(2): p. 144-150.
36.	Yildirim, K., M.L. Hidayetoglu, and A. Capanoglu, Effects of interior colors on mood and preference: comparisons of two living rooms. Perceptual and motor skills, 2011. 112(2): p. 509-524.
37.	Ou, L.C., et al., A cross‐cultural comparison of colour emotion for two‐colour combinations. Color Research & Application, 2012. 37(1): p. 23-43.
38.	Szabo, F., P. Bodrogi, and J. Schanda, Experimental Modeling of Colour Harmony. Color Research and Application, 2010. 35(1): p. 34-49.
39.	He, K.M., et al., Deep Residual Learning for Image Recognition. 2016 Ieee Conference on Computer Vision and Pattern Recognition (Cvpr), 2016: p. 770-778.
40.	Ren, S.Q., et al., Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. Ieee Transactions on Pattern Analysis and Machine Intelligence, 2017. 39(6): p. 1137-1149.
41.	Newell, A., K.U. Yang, and J. Deng, Stacked Hourglass Networks for Human Pose Estimation. Computer Vision - Eccv 2016, Pt Viii, 2016. 9912: p. 483-499.
42.	He, K., et al., Mask r-cnn. arXiv preprint arXiv:1703.06870, 2017.
43.	Canziani, A., A. Paszke, and E. Culurciello, An analysis of deep neural network models for practical applications. arXiv preprint arXiv:1605.07678, 2016.
44.	Ryan, M.-L., Narrative as virtual reality: Immersion and interactivity in literature and electronic media. 2001: Johns Hopkins University Press.
45.	Dengler, P., et al., Scalable vector graphics (SVG) 1.1. 2011, Citeseer.
46.	Krizhevsky, A., I. Sutskever, and G.E. Hinton, ImageNet Classification with Deep Convolutional Neural Networks. Communications of the Acm, 2017. 60(6): p. 84-90.
47.	Szegedy, C., et al. Going deeper with convolutions. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.
48.	Chen, Y., et al., Cascaded pyramid network for multi-person pose estimation. arXiv preprint arXiv:1711.07319, 2017.
49.	Papandreou, G., et al., Towards Accurate Multi-person Pose Estimation in the Wild. 30th Ieee Conference on Computer Vision and Pattern Recognition (Cvpr 2017), 2017: p. 3711-3719.
50.	Yamaguchi, K., et al. Parsing clothing in fashion photographs. in Computer Vision and Pattern Recognition. 2012.
51.	Yamaguchi, K., M.H. Kiapour, and T. Berg, Paper Doll Parsing: Retrieving Similar Styles to Parse Clothing Items. IEEE, 2013: p. 3519-3526.
52.	Xing, L., et al., Intelligent recognition of dominant colors for Chinese traditional costumes based on a mean shift clustering method. The Journal of The Textile Institute, 2018: p. 1-11.
53.	Comaniciu, D. and P. Meer, Mean shift: A robust approach toward feature space analysis. Ieee Transactions on Pattern Analysis And Machine Intelligence, 2002. 24(5): p. 603-619.